Comparing results from a traditional multivariable model and seven propensity score-based models for estimating COVID-19 vaccine effectiveness

Background. Accurate estimation of vaccine effectiveness (VE) in real-world settings is essential for guiding immunization strategies, especially in older populations. However, observational studies are prone to bias due to confounding factors, and the choice of statistical method can significantly influence VE estimates. Materials and methods. We compared the performance of a multivariable Cox proportional hazards model with seven propensity score (PS)-based models to estimate the relative vaccine effectiveness (rVE) of the bivalent Original/Omicron BA.4-5 mRNA vaccine as a second or third booster, compared to a first mRNA booster received ≥120 days earlier. Data from 11,879,461 individuals aged ≥60 in Italy (April-June 2023) were analyzed. Results. All models produced consistent rVE estimates, with values ranging from 16.4% to 22.1%. Over time, booster effectiveness declined, with the reference model showing a drop in rVE from 45.6% (15-60 days) to 14.3% (181-265 days). PS-based methods improved covariate balance but did not outperform the Cox model in terms of precision or interpretability. Conclusions. In large, relatively balanced datasets, traditional multivariable models remain a robust and reliable choice for estimating VE. While PS-based methods offer theoretical advantages, their practical benefit may be limited when confounding is well controlled.